PU-TSI: Interactive Vehicle Trajectory Prediction Considering Perception Information Uncertainty

Abstract

Accurately predicting the future trajectories of surrounding vehicles (SVs) is crucial for enhancing driving safety in Internet of Vehicles (IoV). However, existing trajectory prediction models often suffer from reduced accuracy and stability when confronted with perception information uncertainty. To address this issue, a novel trajectory prediction method considering perception uncertainty (PU-TSI) is proposed, which weakens the propagation of invalid trajectory information, improves prediction accuracy and stability, and adapts to dynamic vehicle interaction scenarios. Specifically, perception uncertainty is modeled based on the Bi-GRU network and integrated into trajectory feature encoding through the proposed data-confidence attention mechanism that jointly accounts for the vehicle motion state and dynamic temporal-spatial interactions. The dynamic interactions are extracted using the graph dual-attention network. In the encoding stage, unlike traditional linear methods, this method fuses different trajectory features considering uncertainty, effectively capturing the complex coupling effects between features and utilizing the full spectrum of available trajectory information. The fused encoded features are passed to the decoding stage, where vehicle interactions guide the trajectory decoder to predict future trajectories. Finally, the proposed model is trained and validated on multiple public datasets. The experiment results demonstrate that PU-TSI effectively predicts the future trajectories of SVs in various interaction scenarios, and achieves superior prediction accuracy and stability compared to existing models.